Classifying GitHub repositories with minimal human efforts

GitHub is a great platform for sharing software code, data, and other resources. To improve search and analysis of a vast spectrum of resources on GitHub, it is necessary to conduct automatic, flexible and user-guided classification of GitHub repositories. In this paper, we study how to build a customized repository classifier with minimal human annotation. Previous document classification methods cannot be directly applied to our task due to three unique challenges: (1) multi-modal signals: besides text, signals in other formats need to be explored to uncover the topic of a repository; (2) low data quality: GitHub README files, usually containing code and commands, are noisier than typical text data such as scientific papers and news; and (3) limited ground-truth: users cannot afford to label many repositories for training a good classifier. To deal with the challenges above, we propose GitClass, a framework to classify GitHub repositories under weak supervision. Three key modules, heterogeneous network construction and embedding, keyword extraction and topic modeling, as well as pseudo document generation, are used to tackle the above three challenges, respectively. We conduct extensive experiments on three large-scale GitHub repository datasets and observe evident performance boost over state-of-the-art embedding and classification algorithms.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2019

Author dc:creator

• Zhang, Yu

Contributors dc:contributor

• Han, Jiawei

• GitHub
• Classification
• Weak Supervision

Statement dc:rights

• Copyright 2019 Yu Zhang

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/104942

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/104942